FPGA-based lightweight target detection neural network implementation method

Abstract

The invention discloses an implementation method of a lightweight target detection neural network based on an FPGA. The network comprises a plurality of convolution layers, a plurality of pooling layers and a frame regression output layer. The method comprises the following basic process steps: constructing a lightweight deep convolutional neural network; initializing a lightweight deep convolutional neural network; training a lightweight deep convolutional neural network; designing a basic part device of a lightweight convolutional neural network in a field programmable gate array FPGA; realizing a trained lightweight convolutional neural network on the FPGA; performing physical testing. According to the invention, on the premise of ensuring the accuracy of the prediction result of the deep convolutional neural network, the structure of the deep convolutional neural network is simplified, the network training method is optimized, the hardware computing unit is optimized, and the resource utilization rate and the operation effect of realizing the deep convolutional neural network in the FPGA are improved.

Description

technical field [0001] The invention relates to a target detection network technology, in particular to an implementation method of a lightweight target detection neural network based on FPGA, and belongs to the technical field of image processing. Background technique [0002] Convolution Neutral Network (hereinafter referred to as CNN) is a general solution in the field of target detection and image classification. The typical deep convolutional neural network has the disadvantages of long training time and high equipment requirements at the same time of high accuracy. However, in practical application scenarios, it needs to be executed in mobile devices and embedded systems with limited computing resources, limited energy, and severe environmental interference. [0003] Therefore, different accelerators based on FPGA, GPU (Graphics Processing Unit) and even ASIC (Application Specific Integrated Circuit) have been proposed to improve the performance of deep network design...

AI Technical Summary

Problems solved by technology

[0005] The target detection based on deep learning has greatly improved the detection accuracy and detection speed compared with the traditional method, but still faces some problems: ①Most of the current algorithms use transfer learning, that is, the existing large data set Carry out training, and then fine-tune (fine-tune) the trained "semi-finished product"

②Deep learning is poorly interpretable, especially at a deeper level. In many cases, it can only rely on tests and experience to guess the reasons for its effectiveness or ineffectiveness. It lacks a clear explanation for the intermediate process, and it is more like a black box.

③High calculation intensity

④The FPGA implementation method of compressed network / pruned network has a great loss of precision and the effect is not ideal; the design of fixed-point strategy usually does not consider the matching of computing throughput and memory bandwidth of the FPGA platform, and does not make full use of logic resources and memory bandwidth. , can not get the best performance

Images